feat: agentic related-code reading (point 1 — biggest quality lever)

[NeverENG]

What

Implements the #1 quality lever from DESIGN.md §六: agentic related-code reading.

Until now BanGD reviewed only the diff plus the full contents of the changed files. It never saw the untouched code a change depends on — the struct/interface defined in another file, the caller that already holds a lock, the package's other methods sharing an invariant. That's exactly the context architecture-level reasoning needs.

How

A new core module src/core/related.ts adds one model-directed planning round before the review:

1. Plan — the model names the untouched files it needs, inferred from the Go imports/identifiers already visible in the changed code (no hallucinated paths).
2. Fetch — the core pulls each via readFile, bounded by count (MAX_RELATED_FILES) and char budget (RELATED_CHAR_BUDGET). A miss returns null and is dropped — cheap.
3. Review — fetched files feed into the review as a dedicated context block.

Deliberate scope (v1)

• One round, reusing the existing single-shot generateStructured port — no multi-turn tool-use redesign. Earns its keep first; multi-round / interleaved tool-use is a documented later refinement.
• Related files land in the user-message tail, so the large cached system prompt (caching) is unaffected.
• Observable: ReviewOutcome.relatedFiles surfaces which files were pulled — shown in the PR comment footer, the Action log, and the smoke output. This is the concrete demonstration of the lever working for a judged deliverable.
• Planning failure is never fatal (returns []; the review proceeds).

Tests

• New src/core/related.test.ts (9 cases): trim/dedupe, exclude changed files, drop non-strings, cap, no-paths/throw → [], schema identity, fetch-and-drop-misses.
• review.test.ts updated to be schema-aware + a new case asserting planner-named files are fetched and injected.
• All 65 tests pass; typecheck + lint clean; dist/ bundle rebuilt and committed.

🤖 Generated with Claude Code